Openflow App Testing Chao SHI, Stephen Duraski Motivation • Network is still a complex stuff ! Distributed mechanism o Complex protocol o Large state o Asynchronous events, random ordering, race conditions. o Motivation • Openflow does not reduce complexity, it just encapsulates it. The distributed and asynchronous feature of networking is inherent o The correctness of controller not verified o Hard to test your apps on Openflow controller o Race Condition Background Finite State Machine • • A model of hardware or computer programs State o • • • Variable value assignment, flags, code block in execution, etc Transition Input (Event) Output (Action) State Diagram Model Checking for Distributed System • • System: o State: o • Components & Channels o Component State System State = Component States + Channel Contents Transition o o Enabled Transitions (System level) System Execution (Sequence of Transitions) Model Checking • What to look for: o • o Method: o • Deadlocks, incompatible states causing crash Both hardware and software o State Space Enumeration Depth First Search using stack in the paper No Scalability Less suitable for software o o Large possibility of user input "Undecidability cannot be fully algorithmic" Symbolic Execution • • • • Tracking the symbolic variable rather than actual value Symbolic variable is associated with input Real variables are associated with symbolic variables and the expression is updated accordingly Fork at every branch. o Still not scalable Symbolic Execution • • • Equivalent class o Same code path for all instances in the same class Still problems o Code path coverage does not mean system path coverage Combine State Modeling and Symbolic Execution is good ! o o State modeling to exhaust system states Symbolic execution to reduce transition numbers Design Modeling in Openflow • • • State: Controller + Switch + Host Transition Input: o • o Event (Packet_IN, Flow_Removed, Port_Status) Message Content (OFMessage) ( We don't want it ) Output: The actual code path of the event Handler (Atomic) o That problem can be solved using the extension of Symbolic Execution o OFSwitch and Host Simplification • • Switch is channels + flow table Switch has Two simple transition o • • o process_pkt, process_of process_pkt: multiple packet processing - > a state transfer Equivalent flow table is the same state Host is categorized o o Server-- receive / send_reply Client -- send / receive Symbolic Execution • • • • Find the equivalent class of input packet Pick one packet from the class as relevant packet Symbolic packet composed of Symbolic variables Concrete Controller State Combination Initialization: enable discover_packets by every switch Run discover_packets_transition and enable new transitions Are you the switch in discover_packets_transiti on. If so, your next transition is already assigned. If not, continue listening on packets state of the ctrl is changed OpenFlow Specific Search Strategies • • • • • PKT-SEQ: Relevant packet sequences NO-DELAY: Instantaneous rule updates UNUSUAL: Uncommon delays and reordings FLOW-IR: Flow independence reduction PKT-SEQ is complementary to other strategies in that it only reduces the number of send transitions rather than the possible kind of event orderings Specifying Application Correctness • • Customizable correctness properties testing correctness involves verifying that "something bad never happens" (safety) and that "something good eventually happens" (liveness). To check these properties, NICE allows correctness properties to (i)access the system state, (ii) register callbacks, and (iii) maintain local state Specifying Application Correctness • • NICE provides a library of correctness properties applicable to a wide variety of OpenFlow applications. Includes: NoForwardingLoops, NoBlackHoles, DirectPaths, StrictDirectPaths, NoForgottenPackets Implementation NICE consists of 3 parts: A model checker - NICE remembers the sequence of transitions that created the state and restores it by replaying the sequence, state-matching is doing by comparing and storing hashes of the explored states. A concolic execution engine - executes the Python code with concrete instead of symbolic inputs A collection of models including the simplified switch and several end hosts. • • • Performance • • • • The experiments are run on the topology of Figure 1 (reproduced here) Host A sends a “layer-2 ping” packet to host B which replies with a packet to A. The controller runs the MAC-learning switch program of Figure 3. The numbers of transitions and unique states, and the execution time as the number of concurrent pings is increased (a pair of packets) are reported. Performance Performance Comparison to state-of-the-art model checkers SPIN and JPF SPIN performs a full search more efficiently, but statespace explosion occurs and SPIN runs out of memory with 7 pings SPIN's Partial-order reduction only decreases the growth rate by 18% JPF is already slower than nice, and even after extra effort is put into improving the Java model used in JPF, it is still substantially slower. • • • NICE and Real Applications • NICE detected 3 bugs in a MAC-learning switch (PySwitch) o Host unreachable after moving o Delayed direct path o Excess flooding NICE and Real Applications • NICE found several bugs in a Web Server Load Balancer (dynamically reconfigures data center traffic based on load) o Next TCP packet always dropped after reconfiguration o Some TCP packets dropped after reconfiguration o ARP packets forgotten during address resolution o Duplicate SYN packets during transitions NICE and Real Applications • Energy-Efficient Traffic Engineering (selectively powering down links and redirecting during light loads) o First packet of new flow is dropped o The first few packets of a new flow can be dropped o Only on-demand routes used under high loads o Packets can be dropped when the load reduces Overhead of running NICE Coverage vs. Overhead Trade-Offs There are some current limitations to NICE with the current approach Concrete execution on the switch Concrete global controller variables Infinite execution trees in symbolic execution • • • Related work • • NICE uses ideas in model checking and symbolic execution, and is the first to apply those techniques to testing OpenFlow networks. Other work such as Frenetic, OFRewind, and FlowChecker can perform limited testing in OpenFlow networks.